The recent refactoring of the dot_prod optab to convert-type exposed a limitation in how `find_widening_optab_handler_and_mode' is currently implemented, owing to the fact that, while the function expects the
GET_MODE_CLASS (from_mode) == GET_MODE_CLASS (to_mode) condition to hold, the c6x backend implements a dot product from V2HI to SI, which triggers an ICE. Consequently, this patch adds some logic to allow widening optabs which accumulate vector elements to a single scalar. Regression tested on x86_64 and aarch64 with no new regressions. Fixes failing unit tests on c6x, as validated for the tic6x-unknown-elf target. Ok for master? gcc/ChangeLog: PR middle-end/116926 * optabs-query.cc (find_widening_optab_handler_and_mode): Add handling of vector -> scalar optab handling. --- gcc/optabs-query.cc | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/gcc/optabs-query.cc b/gcc/optabs-query.cc index c3134d6a2ce..cc52bc0f5ea 100644 --- a/gcc/optabs-query.cc +++ b/gcc/optabs-query.cc @@ -485,6 +485,12 @@ find_widening_optab_handler_and_mode (optab op, machine_mode to_mode, if (GET_MODE_CLASS (limit_mode) == MODE_PARTIAL_INT) limit_mode = GET_MODE_WIDER_MODE (limit_mode).require (); } + else if (is_a <scalar_int_mode> (to_mode)) + { + gcc_checking_assert (VECTOR_MODE_P (from_mode) + && GET_MODE_INNER (from_mode) < to_mode); + limit_mode = from_mode; + } else gcc_checking_assert (GET_MODE_CLASS (from_mode) == GET_MODE_CLASS (to_mode) && from_mode < to_mode); -- 2.34.1